Providing QoS for Real-time Applications - CiteSeerX

Providing QoS for Real-time Applications Teresa Tung, Jean Walrand Department of Electrical Engineering and Computer Sciences University of California, Berkeley, CA 94720 teresat, wlr
@eecs.berkeley.edu

¯

Preserve the ability to serve elastic traffic. Provide fair resource allocation in the sense that neither real-time traffic nor elastic traffic monopolizes network resources. The scheme has the following additional objectives: ¯ Minimize packet loss to make retransmissions less frequent and improve the throughput of the network. ¯ The scheme should admit a simple implementation that does not require complex signaling nor a centralized manager. This paper is organized as follows. Section 2 describes Index Terms—Quality of service, Congestion Control. the scheme. The scheme combines previously proposed ideas in a new way that enables the estimate of its performance. Section 3 provides a theoretical analysis of I. I NTRODUCTION the performance of time-sensitive applications. Section Internet accommodates many types of traffic. These 4 presents simulation results. Section 5 describes an imtypes can be categorized as real-time and elastic. Elasplementation on a test network. Section 6 concludes with tic traffic refers to that of applications where the transmitrelated work and deployment issues. ted information is not time sensitive, but requires eventual correct delivery. Examples of applications that genII. P ROVIDING Q O S TO EF T RAFFIC erate elastic traffic are email, web-browsing, file transfers (FTP), Telnet, and any application that works without This section describes a networking scheme that baltimely delivery. Internet accommodates elastic traffic very ances the needs of applications wtih real-time and elastic well. Protocols like TCP control the transmission rate of traffic. elastic traffic and allow for reliable transmission. Consider VoIP as a representative time-sensitive appliReal-time traffic refers to that of applications where cation. If the network cannot guarantee a small latency the transmitted information is only useful if it is received for a VoIP call, it should block that call. The network can within a small delay. Such traffic often does not benefit block the call with a busy signal as in the public switched from retransmissions of lost packets since they would ar- telephone network. This observation suggests that the netrive too late. Examples of applications that generate real- work should control real-time traffic by some admission time traffic are voice over IP (VoIP), video conferences, control policy based on the latency. and generally any application that requires small end-toIn contrast with real-time traffic, the network can adend delay. Internet does not cater well to the time con- just the transmission of elastic traffic with a congestion straints needed by such real-time applications. There is control policy and does not need to subject such traffic to significant interest in developing an Internet that can ac- admission control. comodate real-time applications. We study the feasibility of a scheme that uses packet This paper studies a scheme to guarantee low end-to- marking based on virtual queues for feedback congestion end delay and loss rate for latency-sensitive traffic without control and admission control. When it gets congested, a reducing the Internet’s ability to transport elastic traffic router marks a packet by turning on a bit in the header of efficiently. The main goals are the following: the packet. The Random Early Detection (RED) scheme ¯ Minimize the queuing delay of time-sensitive flows. is an example of a scheme used to determine the onset of

Abstract—This paper describes a framework for packetswitched networks that balances the needs of real-time and data applications. Routers use virtual queues to mark packets reflecting a measure of congestion. These marks are then used to control the admission of time-sensitive traffic and to regulate other traffic. We present a theoretical analysis of this framework. We evaluate the performance characteristics of this approach in terms of overall link utilization and quality of service of the time-sensitive traffic. We conclude with a discussion of simulation results and of an implementation on a test network.

¯

congestion and to mark packets [8]. Using virtual queues with reduced transmission rates to track the operations of the router with is another method of predicting congestion and marking packets. We study a virtual queue scheme which we describe later. Our scheme subjects elastic traffic to a congestion control scheme that reacts to packet marking. TCP with ECN (Explicit Congestion Notification) is a congestion control scheme ideal for controlling elastic traffic [7]. TCP with ECN reacts to marked packets as if they were dropped packets, by decreasing of the transmission rate of the flow. Using virtual queues, TCP with ECN can decrease a flow’s transmission rate before the router queue builds up. An admission control policy based on packet marking governs the time-sensitive flows. Gibbens and Kelly propose an admission control scheme for time-sensitive traffic where decisions are made in a distributed manner by the source and the destination [10]. We study the same admission control policy as Gibbens and Kelly: the systems we study respond to marked packets in the same manner. However our work differs in how routers mark packets. Gibbens and Kelly study a generalized packet marking scheme. We study a specific scheme using virtual queues for which we can provide more detailed analysis. We investigate the admission control scheme using packet marking based on a specific configuration of virtual queues for which we offer a unique theoretical interpretation, simulation results, and experimental results. These previously proposed schemes have been studied separately using a packet marking system. We study the combination of these congestion control and admission control schemes using a packet marking system based on a pair of virtual queues. Next, we describe the suggested behavior of the router, the source, and the destination.

Real-Time Physical Queue

high Capacity C

B All Packets Elastic Physical Queue

low

B Capacity C 1 Real-Time Virtual Queue C 1 +C2